DGRPool: A web tool leveraging harmonized Drosophila Genetic Reference Panel phenotyping data for the study of complex traits

  1. Laboratory of Systems Biology and Genetics, Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne (EPFL), CH-1015 Lausanne, Switzerland
  2. Swiss Institute of Bioinformatics, CH-1015 Lausanne, Switzerland
  3. Bioinformatics Competence Center, EPFL, Switzerland
  4. KU Leuven, Belgium

Peer review process

Not revised: This Reviewed Preprint includes the authors’ original preprint (without revision), an eLife assessment, public reviews, and a provisional response from the authors.

Read more about eLife’s peer review process.

Editors

  • Reviewing Editor
    Jonathan Flint
    University of California, Los Angeles, Los Angeles, United States of America
  • Senior Editor
    David James
    University of Sydney, Sydney, Australia

Reviewer #1 (Public Review):

This is a technically sound paper focused on a useful resource around the DRGP phenotypes which the authors have curated, pooled, and provided a user-friendly website. This is aimed to be a crowd-sourced resource for this in the future.

The authors should make sure they coordinate as well as possible with the NC datasets and community and broader fly community. It looks reasonable to me but I am not from that community.

I have only one major concern which in a more traditional review setting I would be flagging to the editor to insist the authors did on resubmission. I also have some scene setting and coordination suggestions and some minor textual / analysis considerations.

The major concern is that the authors do not comment on the distribution of the phenotypes; it is assumed it is a continuous metric and well-behaved - broad gaussian. This is likely to be more true of means and medians per line than individual measurements, but not guaranteed, and there could easily be categorical data in the future. The application of ANOVA tests (of the "covariates") is for example fragile for this.

The simplest recommendation is in the interface to ensure there is an inverse normalisation (rank and then project on a gaussian) function, and also to comment on this for the existing phenotypes in the analysis (presumably the authors are happy). An alternative is to offer a kruskal test (almost the same thing) on covariates, but note PLINK will also work most robustly on a normalised dataset.

Minor points:
On the introduction, I think the authors would find the extensive set of human GWAS/PheWAS resources useful; widespread examples include the GWAS Catalog, Open Targets PheWAS, MR-base, and the FinnGen portal. The GWAS Catalog also has summary statistics submission guidelines, and I think where possible meta-data harmonisation should be similar (not a big thing). Of course, DRGP has a very different structure (line and individuals) and of course, raw data can be freely shown, so this is not a one-to-one mapping.

For some authors coming from a human genetics background, they will be interpreting correlations of phenotypes more in the genetic variant space (eg LD score regression), rather than a more straightforward correlation between DRGP lines of different individuals. I would encourage explaining this difference somewhere.

This leads to an interesting point that the inbred nature of the DRGP allows for both traditional genetic approaches and leveraging the inbred replication; there is something about looking at phenotype correlations through both these lenses, but this is for another paper I suspect that this harmonised pool of data can help.

I was surprised the authors did not crunch the number of transcript/gene expression phenotypes and have them in. Is this because this was better done in other datasets? Or too big and annoying on normalisation? I'd explain the rationale to leave these out.

I think 25% FDR is dangerously close to "random chance of being wrong". I'd just redo this section at a higher FDR, even if it makes the results less 'exciting'. This is not the point of the paper anyway.

I didn't buy the extreme line piece as being informative. Something has to be on the top and bottom of the ranks; the phenotypes are an opportunity for collection and probably have known (as you show) and cryptic correlations. I think you don't need this section at all for the paper and worry it gives an idea of "super normals" or "true wild types" which ... I just don't think is helpful.

I'd say "well-established inversion genotypes and symbiot levels" rather than generic covariates. Covariates could mean anything. You have specific "covariates" which might actually be the causal thing.

I wouldn't use the adjective tedious about curation. It's a bit of a value judgement and probably places the role of curation in the wrong way. Time-consuming due to lack of standards and best practice?

Reviewer #2 (Public Review):

Summary:
​In the present study, Gardeux et al provide a web-based tool for curated association mapping results from DRP studies. The tool lets users view association results for phenotypes and compare mean phenotype ~ phenotype correlations between studies. In the manuscript, the authors provide several example utilities associated with this new resource, including pan-study summary statistics for sex, traits, and loci. They highlight cross-trait correlations by comparing studies focused on longevity with phenotypes such as oxphos and activity.

Strengths:
-Considerable efforts were dedicated toward curating the many DRG studies provided.
-Available tools to query large DRP studies are sparse and so new tools present appeal

Weaknesses:
The creation of a tool to query these studies for a more detailed understanding of physiologic outcomes seems underdeveloped. These could be improved by enabling usages such as more comprehensive queries of meta-analyses, molecular information to investigate given genes or pathways, and links to other information such as in mouse rat or human associations.

Author Response

We would like to thank the reviewers for their positive and constructive comments on the manuscript.

We are planning the following revisions to both DGRPool and the corresponding manuscript to address the reviewers’ comments:

  1. We agree with reviewer #1 that normalizing the data could potentially improve the GWAS results. Thus, we plan to explore the implementation of this option and assess its impact on the overall results. We will also investigate replacing the ANOVA test with a KRUSKAL test. Instead of upfront data normalization, we will consider using the PLINK –pheno-quantile-normalize option. Both options will be compared on a set of phenotypes where we can analyze the output (i.e., for phenotypes where we expect to find specific variants), to determine whether these strategies enhance the detection power.

  2. We also agree with both reviewers that gene expression information is of interest. However, we recognize that incorporating such information would entail substantial work (as elaborated in our response to comments below). We feel that this extensive work is beyond the current scope of this paper, which primarily focuses on phenotypes and genotype-phenotype associations. Nonetheless, we are committed to enhancing user experience by including more gene-level outlinks to Flybase. Additionally, we will link variants and gene results to Flybase's online genome browser, JBrowse. By following the reviewers' suggestions, we aim to guide DGRPool users to potentially informative genes.

  3. In agreement with reviewer #2, we acknowledge that additional tools could enhance DGRPool's functionality and facilitate meta-analyses for users. Therefore, we are in the process of developing a gene-centric tool that will allow users to query the database based on gene names. Moreover, we intend to integrate ortholog databases into the GWAS results. This feature will enable users to extend Drosophila gene associations to other species if necessary.

  4. Finally, we also concur with both reviewers about making minor edits to the manuscript to address their feedback.

Reviewer #1 (Public Review):

This is a technically sound paper focused on a useful resource around the DRGP phenotypes which the authors have curated, pooled, and provided a user-friendly website. This is aimed to be a crowd-sourced resource for this in the future.

The authors should make sure they coordinate as well as possible with the NC datasets and community and broader fly community. It looks reasonable to me but I am not from that community.

We thank the reviewer for the positive comments. We are relatively well-connected to the D. melanogaster community and aim to leverage this connection to render the resource as valuable as possible. DGRPool in fact already reflects the input of many potential users and was also inspired by key tools on the DGRP2 website. Furthermore, it also rationalizes why we are often bridging our results with other resources, such as linking out to Flybase, which is the main resource for the Drosophila community at large.

I have only one major concern which in a more traditional review setting I would be flagging to the editor to insist the authors did on resubmission. I also have some scene setting and coordination suggestions and some minor textual / analysis considerations.

The major concern is that the authors do not comment on the distribution of the phenotypes; it is assumed it is a continuous metric and well-behaved - broad gaussian. This is likely to be more true of means and medians per line than individual measurements, but not guaranteed, and there could easily be categorical data in the future. The application of ANOVA tests (of the "covariates") is for example fragile for this.

The simplest recommendation is in the interface to ensure there is an inverse normalisation (rank and then project on a gaussian) function, and also to comment on this for the existing phenotypes in the analysis (presumably the authors are happy). An alternative is to offer a kruskal test (almost the same thing) on covariates, but note PLINK will also work most robustly on a normalised dataset.

We thank the reviewer for raising this interesting point. Indeed, we did not comment on the distribution of individual phenotypes due to the underlying variability from one phenotype to another, as suggested by the reviewer. Some distributions appear normal, while others are clearly not normally distributed. This information is 'visible' to users by clicking on any phenotype; DGRPool automatically displays its global distribution if the values are continuous/quantitative. We acknowledge the reviewer's concerns regarding the use of ANOVA tests. However, we consider it acceptable to perform linear regression (including ANOVA tests) on non-normally distributed data, as only the prediction errors need to follow a normal distribution.

Furthermore, the ANOVA test is solely conducted to assess whether any of the potential covariates (such as well-established inversions and symbiont infection status) are associated with the phenotype of interest. PLINK2 automatically corrects for the effects of these covariates during GWAS by considering them as part of the regression model.

Nevertheless, we concur with the reviewer that normalizing the data could potentially enhance GWAS results. Consequently, we commit to exploring the impact of data normalization on the overall outcomes. Additionally, we will consider replacing the ANOVA test with a KRUSKAL test, and using the PLINK –pheno-quantile-normalize option. We intend to compare both approaches using a set of phenotypes where we can compare the output (i.e., where specific variants are expected to be identified). This comparison will help us determine if either method enhances the detection power.

Minor points:

On the introduction, I think the authors would find the extensive set of human GWAS/PheWAS resources useful; widespread examples include the GWAS Catalog, Open Targets PheWAS, MR-base, and the FinnGen portal. The GWAS Catalog also has summary statistics submission guidelines, and I think where possible meta-data harmonisation should be similar (not a big thing). Of course, DRGP has a very different structure (line and individuals) and of course, raw data can be freely shown, so this is not a one-to-one mapping.

Thank you for the suggestion. We will cite these resources in the Introduction and check the GWAS catalog submission guidelines to compare to the ones we are proposing in this paper.

For some authors coming from a human genetics background, they will be interpreting correlations of phenotypes more in the genetic variant space (eg LD score regression), rather than a more straightforward correlation between DRGP lines of different individuals. I would encourage explaining this difference somewhere.

We appreciate this potential issue and we will make this distinction clearer in the manuscript to avoid any confusion.

This leads to an interesting point that the inbred nature of the DRGP allows for both traditional genetic approaches and leveraging the inbred replication; there is something about looking at phenotype correlations through both these lenses, but this is for another paper I suspect that this harmonised pool of data can help.

We agree with the reviewer and hope that more meta-analyses will be made possible by leveraging the harmonized data that are made available through DGRPool.

I was surprised the authors did not crunch the number of transcript/gene expression phenotypes and have them in. Is this because this was better done in other datasets? Or too big and annoying on normalisation? I'd explain the rationale to leave these out.

This is a very good point raised by the reviewer, and this is in fact something that we initially wanted to do. However, to render the analysis fair and robust, it would require processing all datasets in the same way. This implies cataloging all existing datasets and processing them through the same pipeline. Then, it also requires adding a “cell type” or “tissue” layer, because gene expression data from whole flies is obviously not directly comparable to gene expression data from specific tissues or even specific conditions. This would be key information as phenotypes are often tissue-dependent. So, as implied by the reviewer, we deemed this too big of a challenge beyond the scope of the current paper. Nevertheless, we plan to continue investigating this avenue, especially given the strong transcriptomics background of our lab, in a potential follow-up paper.

I think 25% FDR is dangerously close to "random chance of being wrong". I'd just redo this section at a higher FDR, even if it makes the results less 'exciting'. This is not the point of the paper anyway.

We agree with the reviewer that this threshold implies a higher risk of false positive results. However, this is not an uncommonly used threshold (Li et al., PLoS biology, 2008; Bevers et al., Nature Metabolism, 2019; Hwangbo et al, Elife, 2023), and one that seems robust enough in our analysis since similar phenotypes are significant in different studies. Nevertheless, we will revisit these results and explore how a more stringent threshold may impact the results.

I didn't buy the extreme line piece as being informative. Something has to be on the top and bottom of the ranks; the phenotypes are an opportunity for collection and probably have known (as you show) and cryptic correlations. I think you don't need this section at all for the paper and worry it gives an idea of "super normals" or "true wild types" which ... I just don't think is helpful.

This section of the paper was intended to investigate anecdotal evidence suggesting that certain DGRP lines consistently rank at the top or bottom when examining fitness-related traits. If accurate, this observation could imply that inbreeding might have made these lines generally weaker, potentially introducing bias into studies aimed at uncovering the genetic basis of complex traits. However, as per the analyses presented, we did not discover support for this phenomenon. Nevertheless, we consider this message important to convey. In response to the reviewer's feedback, we intend to provide a clearer explanation of the reasoning behind this section of the paper and its main conclusion.

I'd say "well-established inversion genotypes and symbiot levels" rather than generic covariates. Covariates could mean anything. You have specific "covariates" which might actually be the causal thing.

Thank you. We will update the manuscript accordingly.

I wouldn't use the adjective tedious about curation. It's a bit of a value judgement and probably places the role of curation in the wrong way. Time-consuming due to lack of standards and best practice?

Thank you. We will update the manuscript accordingly.

Reviewer #2 (Public Review):

Summary:

In the present study, Gardeux et al provide a web-based tool for curated association mapping results from DRP studies. The tool lets users view association results for phenotypes and compare mean phenotype ~ phenotype correlations between studies. In the manuscript, the authors provide several example utilities associated with this new resource, including pan-study summary statistics for sex, traits, and loci. They highlight cross-trait correlations by comparing studies focused on longevity with phenotypes such as oxphos and activity.

Strengths:

-Considerable efforts were dedicated toward curating the many DRG studies provided.

-Available tools to query large DRP studies are sparse and so new tools present appeal

Weaknesses:

The creation of a tool to query these studies for a more detailed understanding of physiologic outcomes seems underdeveloped. These could be improved by enabling usages such as more comprehensive queries of meta-analyses, molecular information to investigate given genes or pathways, and links to other information such as in mouse rat or human associations.

We appreciate the reviewer's kind comments.

Regarding the tools, we concur with the reviewer that incorporating additional tools could enhance DGRPool and facilitate users in conducting meta-analyses. Therefore, we intend to introduce a gene-centric tool that enables users to query the database based on gene names. Additionally, we will establish links to ortholog databases within the GWAS results, thereby allowing users to extend fly gene associations to other species, if required.

Furthermore, we have plans to link out to a 'genome browser-like' view (Flybase’s JBrowse tool) of the GWAS results centered around the affected variants/genes. We are considering integrating this feature into the new gene-centric tool as well.

Another potential downstream analysis we are considering is gene-set enrichment. This analysis would involve assessing the enrichment of genes in Gene Ontology or other pathway databases directly from the GWAS results page.

  1. Howard Hughes Medical Institute
  2. Wellcome Trust
  3. Max-Planck-Gesellschaft
  4. Knut and Alice Wallenberg Foundation